> ## Content Index
> Fetch the complete content index at: https://slippage.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Treasury yields hit two-decade highs
- URL: https://slippage.ghost.io/treasury-yields-hit-two-decade-highs/
- Published: 2026-09-26T19:27:56.000Z
- Updated: 2026-09-26T19:27:56.000Z
- Description: The 30-year reached 5.47% and Treasury's own buyback filled only $4.08bn of a $6bn maximum, leaving the long end without a bid it can rely on — a repricing already running through credit, structured product and muni flows.
- Author: Vincent Leung
- Tags: Treasury, Public Credit, Mortgage & ABS, Data & Technology, Municipal, Venues & Exchanges

## Treasury's buyback underfills as the long end breaks 2004 highs

The 30-year Treasury yield rose as much as seven basis points to 5.47% on Thursday, its highest since 2004, and the five-year topped 5% for the first time since 2007, per Bloomberg. The $44bn seven-year auction stopped at 5.085%, the highest since the tenor was reintroduced in 2009, with demand short of expectations. On the same day Treasury's buyback operation for 20- to 30-year paper accepted $4.08bn against a $6bn maximum — the second operation running to come in under the size Treasury set for itself, after the 10 September round took $5.2bn. Swaps now fully price three more quarter-point hikes over the next year with hedging for a fourth. Brij Khurana of Wellington Management put it to Bloomberg: "If they're not going to buy back the max amount after these conditions, when will they?"

*Two editions ago the question was whether tripling the buyback would hold the long end; the answer is that Treasury cannot source the bonds at the price it wants, which makes the facility worthless as a floor for anyone sizing long-duration risk against it. The cost lands in construction rather than execution — duration overlays built around an assumed official bid now carry a basis nobody has marked, and a five-year through 5% resets the discount rate under every private valuation still carried at a 2024 cost of capital.*

[Bloomberg via Asset Securitization Report](https://www.google.com/url?q=https://asreport.americanbanker.com/articles/u-s-30-year-yield-hits-highest-since-2004-as-bond-selloff-deepens&source=gmail&ust=1790461409774000&sa=E) · [Advisor Perspectives](https://www.google.com/url?q=https://www.advisorperspectives.com/articles/2026/09/24/global-bond-rout-highest-30-year-yield&source=gmail&ust=1790461409774000&sa=E)

## SoftBank pays 9.75% while investment grade spreads hold near tights

SoftBank priced $11.1bn of high yield across five tranches this week: $1bn at 3.5 years at 8.625%, $4.5bn at 5.5 years at 9.25%, $4.5bn at 7.5 years at 9.75%, plus €500m at four years at 7.125% and €500m at six years at 8%, according to Reuters. It is the largest high yield sale on record by an Asia-Pacific issuer, and it follows a ¥1trn ($6.3bn) retail bond earlier in the month. Books drew more than $20bn of early interest against an initial $10bn-plus-€1bn launch, Bloomberg reported, with Citigroup leading the dollar tranches and JPMorgan the euros; S&P and Fitch rate the paper BB+. The proceeds go against a $64.6bn commitment to OpenAI and related infrastructure. SoftBank's five-year CDS passed 400bp this week, against roughly 280bp in June and 200bp in October 2025.

*IFR put US investment grade spreads at 78bp over Treasuries and hyperscaler dollar issuance at $220bn for the year in its 18 September edition, which means the fastest-compounding concentration of AI credit risk is being priced almost entirely outside the index most credit mandates are measured against — in high yield, in single-name CDS, and in private structures. Constructing to that benchmark leaves the research process tracking a beta drifting away from the actual exposure, and the only desks that can size it are the ones whose systems put cash and CDS on the same screen, which is a data-model question long before it is a trading one.*

[Reuters via Investing.com](https://www.google.com/url?q=https://www.investing.com/news/stock-market-news/softbank-issues-111-billion-in-bonds-in-openai-financing-push-4914266&source=gmail&ust=1790461409774000&sa=E) · [Bloomberg via Yahoo Finance](https://www.google.com/url?q=https://finance.yahoo.com/markets/stocks/articles/softbank-draws-over-20-billion-084641856.html&source=gmail&ust=1790461409774000&sa=E) · [IFR](https://www.google.com/url?q=https://www.ifre.com/topic-codes/2486617/as-government-bonds-wilt-credit-markets-shine&source=gmail&ust=1790461409774000&sa=E)

## AAA CMBS takes a loss at four-year spread tights

Pimco is facing losses above $35m on the CMBS backing Centre Square in Philadelphia, roughly half of its $58m face position sitting in the top-rated tranche, after the two-tower complex sold for $70m against a 2019 appraisal of $471m; analysts put recovery on the $368m deal at 44 cents, and seven junior tranches are written off entirely. Occupancy had fallen to 28% by June from 93% across 56 tenants when the loan was refinanced in 2020\. Bloomberg reports it as only the third AAA-level CMBS loss since the financial crisis, and Bank of America's Alan Todd said: "You're definitely going to see more of this." The same week, Truist Securities' Brock Wolf told Bloomberg that ABS spreads are at their narrowest in more than four years, with US issuance past $300bn for a third consecutive year and running 11% ahead of 2025 — "it's just a great time to be an issuer" — while a $301m Kansas City data-centre CMBS from Pevco was marketed with triple-As at 225bp, the widest at that level so far, per GlobalCapital.

*Aggregate spreads have compressed to the point where they no longer distinguish between collateral types, and at the same moment the deal level is showing that the triple-A label carries visibly different loss risk depending on what sits beneath it. That converts a ratings bucket into a modelling problem: the senior tranche of a 1974 office building and the senior tranche of a data-centre deal cannot share a risk input any more, most PMS and risk setups still treat them as one line, and the work to separate them is integration and data spend that no index-level spread number will ever signal a need for.*

[Bloomberg via Asset Securitization Report](https://www.google.com/url?q=https://asreport.americanbanker.com/articles/pimco-faces-aaa-debt-loss-after-85-writedown-of-office-complex&source=gmail&ust=1790461409774000&sa=E) · [Asset Securitization Report](https://www.google.com/url?q=https://asreport.americanbanker.com/articles/truist-sees-abs-sales-climbing-15-this-year-as-spreads-narrow&source=gmail&ust=1790461409774000&sa=E) · [GlobalCapital](https://www.google.com/url?q=https://www.globalcapital.com/securitization/article/2gxhge1gwhoncexr3czy8/securitization/cmbs/pevco-markets-301m-data-center-cmbs-at-225bp-for-triple-as&source=gmail&ust=1790461409774000&sa=E)

## The semantic layer, not the model, is what gates AI

Cameron Skinner, global head of solutions engineering for Aladdin, told WatersTechnology on 24 September that BlackRock has spent "a lot of time thinking about" semantics before rolling agents out, framing the open questions as "How do you define a workflow? What is AI-enabled versus what becomes AI-native?" and saying those answers govern how quickly value arrives and how large it is. Three days earlier The TRADE reported Bloomberg Intelligence research finding 75% of respondents expect AI to deliver no more than 10% of workload relief, with European buy-side use concentrated in research, data analysis and note generation and absent from execution. Against that, GIC's Weimin Zhu — who runs Americas fixed income trading across seven to eight markets with six or seven traders, at a fund managing roughly $1.18trn — told Markets Media his desk built internal systems that cut price and liquidity discovery from ten to fifteen minutes down to seconds.

*The distance between a platform still deciding what a workflow is and a desk that has already taken a fifteen-minute task to seconds is not a distance in model quality; it is ownership of the data layer, which GIC has built and most of the platform's clients rent. That sets the real price of an OEMS decision — the licence is the explicit cost, and the semantic and mapping work that has to happen before any agent can act on the data is the total cost of ownership, carried by the client either way and written into no contract. The 75% figure is what that unbilled work looks like from the seat of someone who has not done it.*

[WatersTechnology](https://www.google.com/url?q=https://www.waterstechnology.com/trading-tech/7953273/blackrock-aladdin-prioritizing-semantic-layer-for-agentic-ai&source=gmail&ust=1790461409775000&sa=E) · [The TRADE](https://www.google.com/url?q=https://www.thetradenews.com/tradeplus/european-traders-keeping-ai-at-arms-length-from-execution/&source=gmail&ust=1790461409775000&sa=E) · [Markets Media](https://www.google.com/url?q=https://www.marketsmedia.com/gic-strives-to-build-long-term-trading-partnerships/&source=gmail&ust=1790461409775000&sa=E)

## Record muni ETF inflows are partly tax-loss parking

The iShares National Muni Bond ETF took $1.2bn and the Vanguard Tax-Exempt Bond Index ETF $1.7bn in the week before 23 September, records for both on funds of $46bn and $47bn, Bloomberg reported. Thirty-year muni yields then passed 5% on Thursday for the first time since at least 2011\. Municipal returns are down roughly 1.9% on the year. Vanguard's Nathan Will attributed the flows to valuation, saying investors "may be viewing periods of market volatility as an opportunity to add exposure to a high-quality, tax-advantaged asset class". Chris Brigati of SWBC Investment Services put part of it down to tax-loss harvesting, with the ETFs used "as a temporary way to stay invested in the asset class without having to come up with the ideal swap candidate immediately".

*Last week's edition had this asset class going the other way, with the Schwab muni ETF setting a record redemption; a week later two larger wrappers set record inflows, and by Brigati's account a meaningful part of both numbers is the same money passing through the ETF because the SMA swap candidate was not ready. Flow data that cannot separate conviction from tax mechanics is a poor input to an allocation or distribution decision, and for anyone running muni SMAs at scale it hides how much of the book is mid-harvest and sitting outside its target exposure — a reporting gap that falls on operations, not on the trading desk.*

[Bloomberg via Advisor Perspectives](https://www.google.com/url?q=https://www.advisorperspectives.com/articles/2026/09/23/blackrock-vanguard-muni-etfs-record-inflows-bond-rout&source=gmail&ust=1790461409775000&sa=E) · [Bloomberg](https://www.google.com/url?q=https://www.bloomberg.com/news/articles/2026-09-24/muni-30-year-bond-yield-exceeds-5-for-first-time-since-at-least-2011&source=gmail&ust=1790461409775000&sa=E)

## The CFTC stalls compute futures over a missing benchmark

The CFTC wrote to CME on 21 September extending its review of the exchange's GPU compute futures by 45 days, to 9 November, on the ground that the contracts raise novel or complex issues and that it wants wider industry input. The product, built with Silicon Data on H100 and B200 rental prices, had been announced on 11 August for listing on 5 October. The delay lands beside a WatersTechnology piece of 21 September asking whether the market has a benchmark stable enough to settle against, in which Mizuho's Vishal Gupta is quoted on the absence of one: "That in itself is a problem."

*These are the same story told twice — a regulator and a benchmark question stalling one contract — and the timing is what makes it consequential, because compute-price exposure has already arrived on credit books through data-centre ABS and AI-linked issuance with no listed instrument to offset it. Until it lists there is no settled price series, which makes this a research constraint before an execution one: a hedge ratio cannot be tested against a benchmark that does not yet exist, and the exposure meanwhile sits in portfolios as a basis nobody can mark.*

[GuruFocus](https://www.google.com/url?q=https://www.gurufocus.com/news/9092615/cme-group-delays-launch-of-gpu-futures-amid-regulatory-review&source=gmail&ust=1790461409775000&sa=E) · [WatersTechnology](https://www.google.com/url?q=https://www.waterstechnology.com/emerging-technologies/7953259/gpu-compute-derivatives-are-put-to-the-test&source=gmail&ust=1790461409775000&sa=E) · [CME Group](https://www.google.com/url?q=https://www.cmegroup.com/media-room/press-releases/2026/8/11/cme%5Fgroup%5Fand%5Fsilicondatatolaunchcomputefuturesonoctober5tounloc.html&source=gmail&ust=1790461409775000&sa=E)